Meta announced the Llama 4 lineup with multimodal capabilities

Meta announced the fourth version of the Llama family — a lineup of advanced artificial intelligence models. It includes Llama 4 Scout, Llama 4 Maverick, and Llama 4 Behemoth. All models are trained on large-scale datasets of unlabelled text, images, and video, providing them with broad visual and contextual coverage.

The development of Llama 4 was accelerated following the success of China's DeepSeek, whose R1 and V3 models surpassed previous versions of Llama across several metrics. In response, Meta formed special teams to study competitors' approaches to reducing training costs and scaling.

The Scout and Maverick models are already available for download on Llama.com and through Meta's partners, including Hugging Face. Behemoth remains in the training phase. Meanwhile, the updated Meta AI with Llama 4 has already launched in 40 countries.

According to Meta's internal tests, Maverick outperforms GPT-4o and Gemini 2.0 in programming, multilingual input, and image analysis tasks, but falls short of newer models such as Gemini 2.5 Pro and Claude 3.7 Sonnet.

Scout stands out for its ability to analyze large codebases and summarize documents. Its context window reaches 10 million tokens, enabling it to process massive volumes of visual and textual data. The model can run on a single Nvidia H100 GPU, unlike Maverick, which requires a more powerful DGX system.

Behemoth is the most powerful model in the lineup, with nearly 2 trillion parameters, of which 288 billion are active. It already demonstrates superiority over GPT-4.5, Claude 3.7 Sonnet, and Gemini 2.0 Pro in tasks related to exact sciences, but still trails behind Gemini 2.5 Pro.

Meta has also revised the behavior of Llama 4 models on "sensitive" topics: they now refuse to answer less frequently, providing balanced opinions on political and social issues while maintaining neutrality. The company states that Llama 4 is just the first step in a new phase of development for Meta's open models.

Source: TechCrunch